Fan F (2025)
Publication Language: English
Publication Type: Thesis
Publication year: 2025
URI: https://open.fau.de/handle/openfau/36128
Cone beam computed tomography (CBCT) devices are widely used for both diagnostic and interventional purposes. Due to their mobility, they offer greater flexibility compared to standard computed tomography (CT) devices. Additionally, CBCT systems can be easily integrated with navigation systems, expanding their clinical applications. However, due to the limited detector size, truncation occurs during image acquisition, resulting in missing information. This limitation hinders the use of many conventional algorithms in CBCT applications. With the rise of deep learning (DL) techniques, many CBCT-related challenges can potentially be addressed. Nonetheless, issues such as clinical data scarcity and the selection of suitable models remain. The motivation of this thesis is to leverage synthetic data for training DL models and to identify appropriate models for improved CBCT applications in clinical settings. Within the context of this thesis, image truncation in CBCT leads to distortions of fiducial markers and metallic implants when they are reconstructed outside the field of view (FOV). As a result, conventional marker detection and metal artifact reduction (MAR) algorithms are no longer effective. To address these challenges, DL methods are preferred for both tasks. Additionally, various data simulation pipelines are introduced to generate training datasets, overcoming the issues of data scarcity and manual labeling inaccuracies in clinical cases. Specifically, task-specific data simulation improves marker restoration and ensures accurate marker detection. Moreover, the combination of three-dimensional (3D) projection and two-dimensional (2D) X-ray merging methods enhances MAR performance in clinical CBCT reconstructions. Furthermore, this thesis successfully addresses both metal segmentation and metal inpainting tasks within the MAR framework. The choice of DL models is another key focus of this thesis. Two important conclusions are drawn. First, combining conventional algorithms with DL methods leads to better solutions. For example, the Hough Transform (HT) is integrated with DL methods to provide robust marker detection, while a consistency check algorithm refines segmentation results, ensuring greater consistency across projections. Second, novel network structures incorporating Shifted windows (Swin) Transformers outperform conventional convolutional neural networks (CNNs). In this thesis, Swin Transformer-based networks demonstrate superior performance in both segmentation and image generation tasks. With carefully prepared simulation datasets and trained DL models, clinical imaging problems that could not be addressed by conventional methods can now be effectively tackled. This is evidenced by thorough performance evaluations on clinical datasets. First, all distorted markers in CBCT volumes can be accurately detected. Second, complete metal segmentation is achieved for CBCT scans, facilitating the application of MAR algorithms. Third, improved inpainted CBCT scans are produced. The methodologies developed in this thesis have also proven beneficial in recent research challenges. A Swin Transformer-based network enabled enhanced CT synthesis from CBCT and magnetic resonance imaging (MRI) in the SynthRad2023 challenge. Additionally, a cross-domain MAR algorithm proposed in this work was instrumental in winning the MAR challenge in 2024.
APA:
Fan, F. (2025). Synthetic Data-Driven Deep Learning Approaches for Cone Beam Computed Tomography Applications (Dissertation).
MLA:
Fan, Fuxin. Synthetic Data-Driven Deep Learning Approaches for Cone Beam Computed Tomography Applications. Dissertation, 2025.
BibTeX: Download